Huy's Wiki

reflecting-on-a-year-of-claude-code

2 min read

Reflecting on a Year of Claude Code

Key Points

The Claude Code team looks back one year from launch — when the first demo video got “like two reactions” on Slack and the tool “wasn’t really good.”

  • From one agent to armies. Today: prompting one agent, or an agent prompting agents prompting agents — “a tree of thousands of agents.” Wildly different from a year ago.
  • The single most important idea: every time Claude makes a mistake, don’t just re-prompt it — have it write the lesson to CLAUDE.md or make a skill. Do this and “Claude can just run forever,” getting better instead of repeating mistakes.
  • Verification is the thing everyone misunderstands. It’s not lint / type-check / unit tests (those were always easy to automate). For agents it means: can the agent actually run the thing? That takes mental work to set up and is often not straightforward.
  • Origin of self-verification: with Opus 4, they hooked it up and said “build the feature, then test yourself in bash” — it opened a Claude CLI and tested its own feature. Stunning then; now routine (iOS/Android simulators, desktop computers in the loop).

0 items under this folder.